a r t ic l e s

Cortical tracking of hierarchical linguistic structures

in connected speech
Nai Ding1,2, Lucia Melloni35, Hang Zhang1,68, Xing Tian1,9,10 & David Poeppel1,11
The most critical attribute of human language is its unbounded combinatorial nature: smaller elements can be combined into
larger structures on the basis of a grammatical system, resulting in a hierarchy of linguistic units, such as words, phrases and
sentences. Mentally parsing and representing such structures, however, poses challenges for speech comprehension. In speech,
hierarchical linguistic structures do not have boundaries that are clearly defined by acoustic cues and must therefore be internally
and incrementally constructed during comprehension. We found that, during listening to connected speech, cortical activity of
different timescales concurrently tracked the time course of abstract linguistic structures at different hierarchical levels, such as
2016 Nature America, Inc. All rights reserved.

words, phrases and sentences. Notably, the neural tracking of hierarchical linguistic structures was dissociated from the encoding
of acoustic cues and from the predictability of incoming words. Our results indicate that a hierarchy of neural processing
timescales underlies grammar-based internal construction of hierarchical linguistic structure.

To understand connected speech, listeners must construct a hierarchy RESULTS

of linguistic structures of different sizes, including syllables,
words, phrases and sentences13. It remains puzzling how the brain
simultaneously handles the distinct timescales of the different lin-
guistic structures, for example, from a few hundred milliseconds for
syllables to a few seconds for sentences414. Previous studies have
suggested that cortical activity is synchronized to acoustic features
of speech, approximately at the syllabic rate, providing an initial
timescale for speech processing1519. But how the brain utilizes such
syllabic-level phonological representations closely aligned with the
physical input to build multiple levels of abstract linguistic structure,
and represent these concurrently, is not known. We hypothesized
that cortical dynamics emerge at all timescales required for the
processing of different linguistic levels, including the timescales
npg
Cortical tracking of phrasal and sentential structures
In the first set of experiments, we sought to determine the neural
representation of hierarchical linguistic structure in the absence
of prosodic cues. We constructed hierarchical linguistic structures
using an isochronous, 4-Hz sequence of syllables that were inde-
pendently synthesized (Fig. 1a,b, Supplementary Fig. 1 and
Supplementary Table 1). As a result of the acoustic independence
between syllables (that is, no co-articulation), the linguistic constituent
structure could only be extracted using lexical, syntactic and
semantic knowledge, and not prosodic cues. The materials were first
developed in Mandarin Chinese, in which syllables are relatively
uniform in duration and are also the basic morphological unit
(always morphemes and, in most cases, monosyllabic words).

corresponding to larger linguistic structures such as phrases and
sentences, and that the neural representation of each linguistic level
corresponds to timescales matching the timescales of the respective
linguistic level.
Although linguistic structure building can clearly benefit from
prosodic20,21 or statistical cues22, it can also be achieved purely on
the basis of the listeners grammatical knowledge. To experimentally
isolate the neural representation of the internally constructed hier-
archical linguistic structure, we developed new speech materials
in which the linguistic constituent structure was dissociated from
prosodic or statistical cues. By manipulating the levels of linguistic
abstraction, we found separable neural encoding of each different
linguistic level.
Cortical activity was recorded from native listeners of Mandarin
Chinese using magnetoencephalography (MEG). Given that differ-
ent linguistic levels, that is, the monosyllabic morphemes, phrases
and sentences, were presented at unique and constant rates, the
hypothesized neural tracking of hierarchical linguistic structure
was tagged at distinct frequencies.
The MEG response was analyzed in the frequency domain and
we extracted response power in every frequency bin using an opti-
mal spatial filter (Online Methods). Consistent with our hypothesis,
the response spectrum showed three peaks at the syllabic rate (P = 1.4
105, paired one-sided t test, false discovery rate (FDR) corrected),
phrasal rate (P = 1.6 104, paired one-sided t test, FDR corrected)
and sentential rate (P = 9.6 107, paired one-sided t test, FDR

1Department of Psychology, New York University, New York, New York, USA. 2College of Biomedical Engineering and Instrument Sciences, Zhejiang University,
Hangzhou, China. 3Department of Neurology, New York University Langone Medical Center, New York, New York, USA. 4Department of Neurophysiology, Max-Planck
Institute for Brain Research, Frankfurt, Germany. 5Department of Psychiatry, Columbia University, New York, New York, USA. 6Department of Psychology and Beijing
Key Laboratory of Behavior and Mental Health, Peking University, Beijing, China. 7PKU-IDG/McGovern Institute for Brain Research, Peking University, Beijing, China.
8Peking-Tsinghua Center for Life Sciences, Beijing, China. 9New York University Shanghai, Shanghai, China. 10NYU-ECNU Institute of Brain and Cognitive Science

at NYU Shanghai, Shanghai, China. 11Neuroscience Department, Max-Planck Institute for Empirical Aesthetics, Frankfurt, Germany. Correspondence should be
addressed to N.D. (ding_nai@zju.edu.cn) or D.P. (david.poeppel@nyu.edu).

Received 12 August; accepted 3 November; published online 7 December 2015; doi:10.1038/nn.4186

158 VOLUME 19 | NUMBER 1 | JANUARY 2016 nature NEUROSCIENCE

 a r t ic l e s

Figure 1 Neural tracking of hierarchical linguistic structures. a Sentence Sentence 1 Hz

(a) Sequences of Chinese or English monosyllabic words were presented
isochronously, forming phrases and sentences. (b) Spectrum of stimulus N phrase V phrase N phrase V phrase 2 Hz

intensity fluctuation revealed syllabic rhythm, but no phrasal or sentential

4 Hz
modulation. The shaded area covers 2 s.e.m. across stimuli. (c) MEG-
derived cortical response spectrum for Chinese listeners and materials
Dry Fur Rubs Skin New Plans Gave Hope
(dark red curve, grand average; light red curves, individual listeners;
N = 16, 0.11-Hz frequency resolution). Neural tracking of syllabic, 250 ms
phrasal and sentential rhythms was reflected by spectral peaks at
corresponding frequencies. Frequency bins with significantly stronger b Spectrum for stimulus intensity
*
power than neighbors (0.5 Hz range) are marked (*P < 0.001, paired 20 dB

Intensity
one-sided t test, FDR corrected). The topographical maps of response
power across sensors are shown for the peak frequencies.

1 Hz 2 Hz 3 Hz 4 Hz
corrected) and the response was highly consistent across listeners Neural response spectrum
(Fig. 1c). Given that the phrasal- and sentential-rate rhythms were
not conveyed by acoustic fluctuations at the corresponding frequen- c fsentence fphrase fsyllable

cies (Fig. 1b), cortical responses at the phrasal and sentential rates * * *
must be a consequence of internal online structure building processes.

Power (dB)
Cortical activity at all the three peak frequencies was seen bilater- Max
ally (Fig. 1c). The response power averaged over sensors in each
hemisphere was significantly stronger in the left hemisphere at the
2016 Nature America, Inc. All rights reserved.

sentential rate (P = 0.014, paired two-sided t test), but not at the
phrasal (P = 0.20, paired two-sided t test) or syllabic rates (P = 0.40,
paired two-sided t test).
Dependence on syntactic structures
Are the responses at the phrasal and sentential rates indeed separate
neural indices of processing at distinct linguistic levels or are they
merely sub-harmonics of the syllabic rate response, generated by
intrinsic cortical dynamical properties? We address this question by
manipulating different levels of linguistic structure in the input. When
the stimulus is a sequence of random syllables that preserves the
acoustic properties of Chinese sentences (Fig. 1 and Supplementary
Fig. 2), but eliminates the phrasal/sentential structure, only syllabic
(acoustic) level tracking occurs (P = 1.1 104 at 4 Hz, paired one-
sided t test, FDR corrected; Fig. 2a). Furthermore, this manipulation
preserves the position of each syllable in a sentence (Online Methods)
and therefore further demonstrates that the phrasal- and sentential-
rate responses are not a result of possible acoustic differences between
npg
6 dB
1 Hz 2 Hz 3 Hz 4 Hz
Min
Frequency
To test whether the phrase-level responses segregate from the sen-
tence level, we constructed longer verb phrases that were unevenly
divided into a monosyllabic verb followed by a three-syllable noun
phrase (Fig. 2c). We expect that the neural responses to the long
verb phrase to be tagged at 1 Hz, whereas the neural responses to the
monosyllabic verb and the three-syllable noun phrase will present as
harmonics of 1 Hz. Consistent with our hypothesis, cortical dynam-
ics emerged at one-fourth of the syllabic rate, whereas the response
at half of the syllabic rate is no longer detectable (P = 1.9 104, 1.7
104 and 9.3 104 at 1, 3 and 4 Hz, respectively, paired one-sided
t test, FDR corrected).
Dependence on language comprehension
When listening to Chinese sentences (Fig. 1a), listeners who did not
understand Chinese only showed responses to the syllabic (acoustic)

the syllables in a sentence. When two adjacent syllables and mor- rhythm (P = 3.0 105 at 4 Hz, paired one-sided t test, FDR corrected;
phemes combine into verb phrases, but there is no four-element sen- Fig. 2d), further supporting the argument that cortical responses
tential structure, phrasal-level tracking emerges at half of the syllabic to larger, abstract linguistic structures is a direct consequence of
rate (P = 8.6 104 at 2 Hz and P = 2.7 104 at 4 Hz, paired one-sided language comprehension.
t test, FDR corrected; Fig. 2b). Similar responses are observed for If aligning cortical dynamics to the time course of linguistic constit-
noun phrases (Supplementary Fig. 3). uent structure is a general mechanism required for comprehension,
it must apply across languages. Indeed, when native English speakers
Chinese materials, Chinese listeners Chinese materials, English listeners were tested with English materials (Fig. 1a), their cortical activity also
a *
d * followed the time course of larger linguistic structures, that is, phrases
12 dB 12 dB and sentences (P = 4.1 105, syllabic rate; Fig. 2e; P = 3.9 103,
Power

Power

1 2 3 4
1 2 3 4 Figure 2 Tracking of different linguistic structures. Each panel shows
b * * e English materials, English listeners syntactic structure repeating in the stimulus (left) and the cortical
response spectrum (right; shaded area indicates 2 s.e.m. over listeners,
Power

*
N = 8). (a) Chinese listeners, Chinese materials: syllables were
Power

syntactically independent and cortical activity encoded only acoustic and

1 2 3 4
syllabic rhythm. (b,c) Additional tracking emerged with larger linguistic
c * f * * structures. Spectral peaks marked by a star (black, P < 0.001; gray,
* * P < 0.005; paired one-sided t test, FDR corrected). (d) English listeners,
Power
Power

* Chinese materials from Figure 1: acoustic tracking only, as there was

no parsable structure. (e,f) English listeners, English materials: syllabic
1 2 3 4 1/1.28 2/1.28 4/1.28 rate (4/1.28 Hz) and sentential and phrasal rate responses to parsable
Frequency (Hz) Frequency (Hz) structure in stimulus.

nature NEUROSCIENCE VOLUME 19 | NUMBER 1 | JANUARY 2016 159

 a r t ic l e s

Figure 3 Dissociating sentential structures and transitional probability. a Constant transitional probability c Transitional Fourier
(a,b) Grammar of an artificial Markovian stimulus set with constant (a) 1/5 1/5 1/5 probability spectrum
C1 C2 C3 fc
or variable (b) transitional probability. Each sentence consists of three 1/5
acoustic chunks, each containing 12 English words. The listeners
the boy ordered beer, soup, salad C1 C2 C3 C1 C2 C3
memorized the grammar before experiments. (c) Schematic time course a girl lives in pizza, coffee 1 1 1 1
.. fs fc
and spectrum of the transitional probability. (d) Neural response spectrum her dad speaks . 1/25 1/25
(shaded area covers 2 s.e.m. over listeners, N = 8). Significant neural John didnt book, letter, story
Jess wrote a poem, memo Time Frequency
responses to sentences were seen for both languages. Spectral peaks are
shown by an asterisk (P < 0.001, paired one-sided t test, FDR corrected,
b d Constant transitional probability
Varying transitional probability
same color code as the spectrum). Responses were not significantly Varying transitional probability
different between the two languages in any frequency bin (paired ** fs **
1/25 1 1 fc
two-sided t test, P > 0.09, uncorrected). C1 C2 C3 6 dB
**

Power (dB)
my cat is so lovely
4.3 103 and 6.8 106 at the sentential, phrasal and syllabic rates, N = 25 they grow apples
Sarah looks happy
respectively; Fig. 2f; paired one-sided t test, FDR corrected). ... ... ... 1/1.05 2/1.05 3/1.05
Frequency (Hz)

Neural tracking of linguistic structures rather than probability cues

We found that concurrent neural tracking of multiple levels of lin- 2.7 106, respectively). For the varying transitional probability MSS,
guistic structure was not confounded with the encoding of acoustic the response was statistically significant at the sentential rate, twice
cues (Figs. 1 and 2). However, is this simply explained by the neural the sentential rate and the syllable rate (P = 7.1 104, 7.1 104 and
tracking of the predictability of smaller units? As a larger linguistic 4.8 106, respectively).
structure, such as a sentence, unfolds in time, its component units Given that the MSS involved real English sentences, listeners had
2016 Nature America, Inc. All rights reserved.

become more predictable. Thus, cortical networks solely tracking prior knowledge of the transitional probabilities between acoustic
transitional probabilities across smaller units could show temporal chunks. To control for the effect of such prior knowledge, we created
dynamics matching the timescale of larger structures. To test this a set of Artificial Markovian Sentences (AMS). In the AMS, the tran-
alternative hypothesis, we crafted a constant transitional probability sitional probability between syllables was the same in and across sen-
Markovian Sentence Set (MSS) in which the transitional probability tences (Supplementary Fig. 4a). The AMS was composed of Chinese
of lower level units was dissociated from the higher level structures syllables, but no meaningful Chinese expressions were embedded in
(Fig. 3a and Supplementary Fig. 1e,f). The constant transitional the AMS sequences. As the AMS was not based on the grammar of
probability MSS is contrasted with a varying transitional probability Chinese, the listeners had to learn the AMS grammar to segment
MSS, in which the transitional probability is low across sentential sentences. By comparing the neural responses to the AMS sequences
boundaries and high in a sentence (Fig. 3b,c). If cortical activity only before and after the grammar was learned, we were able to separate
encodes the transitional probability between lower level units (for the effect of prior knowledge of transitional probability and the effect
example, acoustic chunks in the MSS) independent of the underlying of grammar learning. Here, the grammar of the AMS indicates the
syntactic structure, it can show tracking of the sentential structure for set of rules that governs the sequencing of the AMS, that is, the rule
the varying probability MSS, but not for the constant probability MSS. of which syllables can follow which syllables.
In contrast with this prediction, indistinguishable neural responses The neural responses to the AMS before and after grammar learning
to sentences were observed for both MSS (Fig. 3d), demonstrating were analyzed separately (Supplementary Fig. 4). Before learning,
that neural tracking of sentences is not confounded by transitional when the listeners were instructed that the stimulus was just a sequence
probability. Specifically, for the constant transitional probability MSS, of random syllables, the response showed a statistically significant
npg

the response was statistically significant at the sentential rate, twice peak at the syllabic rate (P = 0.0003, bootstrap), but not at the senten-
the sentential rate and the syllable rate (P = 1.8 104, 2.3 104 and tial rate. After the AMS grammar was learned, however, a significant
response peak emerged at the sentential rate (P = 0.0001, bootstrap).
A response peak was also observed at twice the sentential rate, possibly
a Neural tracking of sentences
of variable durations (48 syllables)
c Single-trial decoding
of sentence duration reflecting the second harmonic of the sentential response. This result
4 40 further confirms that neural tracking of sentences is not confounded
(number of syllables)

Percent trials (%)

Decoded duration

5 by neural tracking of transitional probability.

RMS over sensors

30
6
6 dB 7
20 Neural tracking of sentences varying in duration and structure
Sentence 10
These results are based on sequences of sentences that have uniform
4 5 6 8
duration:
7 8
duration and syntactic structure. We next addressed whether cortical
4 5 6 7 8
2.25 s Actual duration (number of syllables)

b d Figure 4 Neural tracking of sentences of varying structures. (a) Neural

6 dB activity tracked the sentence duration, even when the sentence boundaries
(dotted lines) were not separated by acoustic gaps. (b) Averaged response
RMS over sensors

near a sentential boundary (dotted line). The power continuously changed

RMS over sensors

throughout the duration of a sentence. Shaded area covers 2 s.e.m.

6 dB over listeners (N = 8). Significance power differences between time bins
(shaded squares) are marked by asterisks (P = 0.01, one-sided t test,
FDR corrected). (c) Confusion matrix for neural decoding of the sentence
* 3-syllable NP VP
* * * * * duration. (d) Neural activity tracks noun phrase duration (shown in
5 4 3 2 1 1 2 4-syllable NP VP the bottom). Yellow areas show significant differences between curves
1.25 s 1s (P = 0.005, bootstrap, FDR corrected).

160 VOLUME 19 | NUMBER 1 | JANUARY 2016 nature NEUROSCIENCE


Figure 5 Localizing cortical sources of the sentential and phrasal rate
responses using ECoG (N = 5). Left, power envelope of high-gamma
activity. Right, waveform of low-frequency activity. Electrodes in the right
hemisphere were projected to the left hemisphere, and right hemisphere
(left hemisphere) electrodes are shown by hollow (filled) circles.
The figure only displays electrodes that showed statistically significant
neural responses to sentences in Figure 2e and no significant response to
the acoustic control shown in Figure 2f. Significance was determined by
bootstrap (FDR corrected) and the significance level is 0.05. The response
strength, that is, the response at the target frequency relative to the mean
response averaged over a 1-Hz wide neighboring region, is color coded.
Electrodes with response strength less than 10 dB are shown by smaller
symbols. The sentential and phrasal rate responses were seen in bilateral
pSTG, TPJ and left IFG.
tracking of larger linguistic structures generalizes to sentences that
are variable in duration (48 syllables) and syntactic structures.
These sentences were again built on isochronous Chinese syllables,
intermixed and sequentially presented without any acoustic gap at
the sentence boundaries. Examples translated into English include
Dont be nervous, The book is hard to read, and Over the street
is a museum.
As these sentences have irregular durations that are not tagged by
2016 Nature America, Inc. All rights reserved.
frequency, the MEG responses were analyzed in the time domain by
averaging sentences of the same duration. To focus on sentential level
processing, we low-pass filtered the response at 3.5 Hz. The MEG
response (root mean square, r.m.s., over all sensors) rapidly increased
after a sentence boundary and continuously changed throughout the
duration of a sentence (Fig. 4a). To illustrate the detailed temporal
a High-gamma power b Low-frequency waveform
fs fp f fs fp
6 dB N = 22
N = 10
npg
N = 13
1/1.28 2/1.28 4/1.28 1/1.28 2/1.28
Frequency (Hz) Frequency (Hz)
fs+ fp fs fp+ (fs+
c d
nature NEUROSCIENCE VOLUME 19 | NUMBER 1 | JANUARY 2016 161
a r t ic l e s
High-gamma power Low-frequency waveform
16 dB
Sentential rate
10 dB
Left hemisphere
Right hemisphere
16 dB
Phrasal rate
10 dB
dynamics, we averaged the r.m.s. response over all sentences contain-
ing six or more syllables after aligning them to the sentence offset
(Fig. 4b). During the last four syllables of a sentence, the r.m.s.
response continuously and significantly decreased for every syllable,
indicating that the neural response continuously changes during the
course of a sentence rather than being a transient response only occur-
ring at the sentence boundary.
A single-trial decoding analysis was performed to independ-
ently confirm that cortical activity tracks the duration of sentences
(Fig. 4c). The decoder applied template matching for the response
time course (leave-one-out cross-validation, Online Methods) and
correctly determined the duration of 34.9 0.6% sentences (mean
s.e.m. over subjects, significantly above chance, P = 1.3 107,
one-sided t test).
f After demonstrating cortical tracking of sentences, we further
tested whether cortical activity also tracks the phrasal structure inside
N = 62 of a sentence. We constructed sentences that consist of a noun phrase
followed by a verb phrase and manipulated the duration of the noun
phrase (three syllable or four syllable). The cortical responses closely
follow the duration of the noun phrase: the r.m.s. response gradually
N = 60 decreased in the noun phrase, then showed a transient increase after
the onset of the verb phrase (Fig. 4d).
Neural source localization using electrocorticography (ECoG)
N = 29
We found that large-scale neural activity measured by MEG concur-
rently follows the hierarchical linguistic structure of speech, but which
neural networks generate such activity? To address this question,
4/1.28 we recorded the ECoG responses to English sentences (Fig. 2e) and
an acoustic control (Fig. 2f). ECoG signals are mesoscopic neuro-
fp+) f
physiological signals recorded by intracranial electrodes implanted in
Figure 6 Spatial dissociation between sentential-rate, phrasal-rate and
syllabic-rate responses (N = 5). (a) The power spectrum of the power
envelope of high-gamma activity. (b) The power spectrum of low-frequency
ECoG waveform. The top panels (green curves) show the response
averaged over all electrodes that show a significant sentential-rate
response but not a significant phrasal-rate response. Significance was
determined by bootstrap (FDR corrected) and the significance level is
0.05. The shaded area is 1 s.d. over electrodes on each side.
The blue curves show the response averaged over all electrodes that
showed a significant phrasal-rate response, but not a significant
sentential-rate response. The red curves show a significant sentential-
rate or a significant phrasal-rate response, but not a significant syllabic
response. (c,d) The topographic distribution of the three groups of
electrodes analyzed in a and b. As in Figure 5, electrodes showing a
response greater than 10 dB are shown by larger symbols than electrodes
showing a response weaker than 10 dB.
 a r t ic l e s
Figure 7 Syllabic-rate ECoG responses to English sentences and the
acoustic control (N = 5). Top, electrodes showing statistically significant
syllabic-rate ECoG responses to the acoustic control, that is, shuffled
sequences, which had the same acoustic and syllabic rhythm as the
English sentences, but contained no hierarchical linguistic structures
(Fig. 2f). Significance was determined by bootstrap (FDR corrected) and
the significance level is 0.05. The responses were most strongly seen
in bilateral STG for both high-gamma and low-frequency activity and in
bilateral pre-motor areas for low-frequency activity. Bottom, syllabic-rate
ECoG responses to English sentences. The electrodes displayed are those
that showed statistically significant neural responses to sentences and no
significant response to the acoustic control. The syllabic rate responses
specific to sentences were strong along bilateral STG for high-gamma
activity and were widely distributed in the frontal and temporal lobes for
low-frequency activity.
epilepsy patients for clinical evaluation (see Supplementary Fig. 5
for the electrode coverage), and they possess better spatial resolution
than MEG. We first analyzed the power of the ECoG signal in the
high gamma band (70200 Hz), as it highly correlates with multiunit
firing23. The electrodes exhibiting significant sentential, phrasal and
syllabic rate fluctuations in high gamma power are shown separately
2016 Nature America, Inc. All rights reserved.
(Fig. 5). The sentential rate response clustered over the posterior and
middle superior temporal gyrus (pSTG), bilaterally, with a second
cluster over the left inferior frontal gyrus (IFG). Phrasal rate responses
were also observed over the pSTG bilaterally. Notably, although the
sentential and phrasal rate responses were observed in similar cortical
areas, electrodes showing phrasal rate responses only partially over-
lapped with electrodes showing sentential rate responses in the pSTG
(Fig. 6). For electrodes showing a significant response at either the
sentential rate or the phrasal rate, the strength of the sentential rate
response was negatively correlated with the strength of the phrasal
rate response (R = 0.32, P = 0.004, bootstrap). This phenomenon
demonstrates spatially dissociable neural tracking of the sentential
and phrasal structures.
Furthermore, some electrodes with a significant sentential or
phrasal rate response showed no significant syllabic rate response
(P < 0.05, FDR corrected, Fig. 6). In other words, there are cortical
circuits specifically encoding larger, abstract linguistic structures
without responding to syllabic-level acoustic features of speech.
npg
In addition, although the syllabic responses were not significantly
different (P > 0.05, FDR corrected) for English sentences and the
acoustic control in the MEG results, they were dissociable spatially
in the ECoG results (Fig. 7). Electrodes showing significant syllabic
responses (P < 0.05, FDR corrected) to sentences, but not the acoustic
control, were seen in bilateral pSTG, bilateral anterior STG (aSTG),
and left IFG.
We then analyzed neural tracking of the sentential, phrasal and
syllabic rhythms in the low-frequency ECoG waveform (Fig. 5), which
is a close neural correlate of MEG activity. Fourier analysis was directly
applied to the ECoG waveform and the Fourier coefficients at 1, 2 and
4 Hz are extracted. Low-frequency ECoG activity is usually viewed as
the dendritic input to a cortical area24. The low-frequency responses
are more distributed than high-gamma activity, possibly reflecting
the fact that the neural representations of different levels of linguistic
structures serve as inputs to broad cortical areas. Sentential and
phrasal rate responses are strong in STG, IFG and temporoparietal
junction (TPJ). Compared with the acoustic control, the syllabic-rate
response to sentences was stronger in broad cortical areas, including
the temporal and frontal lobes (Fig. 7). Similar to the high-gamma
activity, the low-frequency responses to the sentential and phrasal
structures were not reflected in the same set of electrodes (Fig. 6).
162 VOLUME 19 | NUMBER 1 | JANUARY 2016 nature NEUROSCIENCE
High-gamma power Low-frequency waveform
28 dB
Acoustic control
10 dB
Left hemisphere
Right hemisphere
16 dB
Sentence specific
10 dB
For electrodes showing a significant response at either the sentential
rate or the phrasal rate, the strength of the sentential rate response was
also negatively correlated with the strength of the phrasal rate response
(R = 0.21, significantly greater than 0, P = 0.023, bootstrap).
DISCUSSION
Our data show that the multiple timescales that are required for the
processing of linguistic structures of different sizes emerge in corti-
cal networks during speech comprehension. The neural sources for
sentential, phrasal and syllabic rate responses are highly distributed
and include cortical areas that have been found to be critical for
prosodic (for example, right STG), syntactic and semantic (for exam-
ple, left pSTG and left IFG) processing9,2528. Neural integration on
different timescales is likely to underlie the transformation from
shorter lived neural representations of smaller linguistic units to longer
lasting neural representations of larger linguistic structures1114.
These results underscore the undeniable existence of hierarchical
structure building operations in language comprehension1,2 and can
be applied to objectively assess language processing in children and
difficult-to-test populations, as well as animal preparations to allow
for cross-species comparisons.
Relation to language comprehension
Concurrent neural tracking of hierarchical linguistic structures
provides a plausible functional mechanism for temporally integrat-
ing smaller linguistic units into larger structures. In this form of
concurrent neural tracking, the neural representation of smaller
linguistic units is embedded at different phases of the neural activity
tracking a higher level structure. Thus, it provides a possible
mechanism to transform the hierarchical embedding of linguistic
structures into hierarchical embedding of neural dynamics, which
may facilitate information integration in time 10,11. Low-frequency
neural tracking of linguistic structures may further modulate higher
frequency neural oscillations2931, which have been proposed to pro-
vide additional roles for structure building7. In addition, multiple
resources and computations are needed for syntactic analysis, for
example, access to combinatorial syntactic subroutines, and such
operations may correspond to neural computations on distinct
frequency scales, which are coordinated by the low-frequency
neural tracking of linguistic constituent structures. Furthermore,
low-frequency neural activity and oscillations have been hypoth-
esized as critical mechanisms to generate predictions about future
events32. For language processing, it is likely that concurrent neural
tracking of hierarchical linguistic structures provides mechanisms
to generate predictions on multiple linguistic levels and allow
interactions across linguistic levels33.

Neural entrainment to speech
Recent work has shown that cortex tracks the slow acoustic fluctuations
of speech below 10 Hz (refs. 1518,34,35), and this phenomenon is
commonly described as cortical entrainment to the syllabic rhythm of
speech. It has been controversial whether such syllabic-level cortical
entrainment is related to low-level auditory encoding or language
processing6. Our findings demonstrate that processing goes well
beyond stimulus-bound analysis: cortical activity is entrained to larger
linguistic structures that are, by necessity, internally constructed,
based on syntax. The emergence of slow cortical dynamics provides
timescales suitable for the analysis of larger chunk sizes13,14.
A long-lasting controversy concerns how the neural responses to
sensory stimuli are related to intrinsic, ongoing neural oscillations.
This question is heavily debated for the neural response entrained
to the syllabic rhythm of speech36 and can also be asked for neural
activity entrained to the time courses of larger linguistic structures.
Our experiment was not designed to answer this question; however,
we clearly found that cortical speech processing networks have the
capacity to generate activity on very long timescales corresponding
to larger linguistic structures, such as phrases and sentences. In other
words, the timescales of larger linguistic structures fall in the
2016 Nature America, Inc. All rights reserved.
timescales, or temporal receptive windows12,13, that the relevant
cortical networks are sensitive to. Whether the capacity of generating
low-frequency activity during speech processing is the same as the
mechanisms generating low-frequency spontaneous neural oscilla-
tions will need to be addressed in the future.
Nature of the linguistic representations
Language processing is not monolithic and involves partially
segregated cortical networks for the processing of, for example, pho-
nological, syntactic and semantic information9. The phonological,
syntactic and semantic representations are all hierarchically
organized37 and may rely on the same core structure building opera-
tions38. In natural speech, linguistic structure building can be facili-
tated by prosodic39 and statistical cues22, and some underlying neural
signatures have been demonstrated6,8,20. Such cues, however, are not
always available, and even when they are available, they are generally
not sufficient. Thus, robust structure building relies on a listeners
tacit syntactic knowledge, and our findings provide unique insights
npg
into the neural representation of abstract linguistic structures that
are internally constructed on the basis of syntax alone. Although the
construction of abstract structures is driven by syntactic analysis, when
such structures are built, different aspects of the structure, including
semantic information, can be integrated in the neural representa-
tion. Indeed, the wide distribution of cortical tracking of hierarchical
linguistic structures suggests that it is a general neurophysiological
mechanism for combinatorial operations involved in hierarchical
linguistic structure building in multiple linguistic processing networks
(for example, phonological, syntactic and semantic). Furthermore,
coherent synchronization to the correlated linguistic structures in dif-
ferent representational networks, for example, syntactic, semantic and
phonological, provides a way to integrate multi-dimensional linguistic
representations into a coherent language percept38,40, just as tempo-
ral synchronization between cortical networks provides a possible
solution to the binding problem in sensory processing41.
Relation to event-related responses
Although many previous neurophysiological studies on structure
building have focused on syntactic and semantic violations4244, fewer
have addressed normal structure building; on the lexical-semantic
level, the N400/N400m has been identified as a marker of the semantic
nature NEUROSCIENCE VOLUME 19 | NUMBER 1 | JANUARY 2016 163
a r t ic l e s
predictability of words43,45 and its amplitude continuously reduces in
a sentence46,47. For syntactic processing, when two words combine
into a short phrase, increased activity is seen in the temporal and
frontal lobes4. Our results build on and extends these findings by
demonstrating structure building at different levels of the linguis-
tic hierarchy, during online comprehension of connected speech
materials in which the structural boundaries are neither physically
cued nor confounded by the semantic predictability of the individual
words (Fig. 3). Note that, although the two Markovian languages
(compared in Fig. 3) differed in their transitional probability between
acoustic chunks, they both had fully predictable syntactic structures.
The equivalence in syntactic predictability is likely to result in the
very similar responses between the two conditions.
Lastly, the emergence of slow neural dynamics tracking superor-
dinate stimulus structures is reminiscent of what has been observed
during decision making48, action planning49 and music perception50,
suggesting a plausible common neural computational framework to
integrate information over distinct timescales12. These findings invite
MEG and EEG studies to extend from the classic event-related designs
to investigating continuous neural encoding of internally constructed
perceptual organization of an information stream.
Methods
Methods and any associated references are available in the online
version of the paper.
Note: Any Supplementary Information and Source Data files are available in the
online version of the paper.
Acknowledgments
We thank J. Walker for MEG technical support, T. Thesen, W. Doyle and
O. Devinsky for their instrumental help in collecting ECoG data, and G. Buzsaki,
G. Cogan, S. Dehaene, A.-L. Giraud, G. Hickok, N. Hornstein, E. Lau, A. Marantz,
N. Mesgarani, M. Pea, B. Pesaran, L. Pylkknen, C. Schroeder, J. Simon and
W. Singer for their comments on previous versions of the manuscript. This work
supported by US National Institutes of Health grant 2R01DC05660 (D.P.) and
Major Projects Program of the Shanghai Municipal Science and Technology
Commission (STCSM) 15JC1400104 (X.T.) and National Natural Science
Foundation of China 31500914 (X.T.).
AUTHOR CONTRIBUTIONS
N.D., L.M. and D.P. conceived and designed the experiments. N.D., H.Z. and X.T.
performed the MEG experiments. L.M. performed the ECoG experiment.
N.D., L.M. and D.P. wrote the paper. All of the authors discussed the results
and edited the manuscript.
COMPETING FINANCIAL INTERESTS
The authors declare no competing financial interests.
Reprints and permissions information is available online at http://www.nature.com/
reprints/index.html.
1. Berwick, R.C., Friederici, A.D., Chomsky, N. & Bolhuis, J.J. Evolution, brain, and
the nature of language. Trends Cogn. Sci. 17, 8998 (2013).
2. Chomsky, N. Syntactic Structures (Mouton de Gruyter, 1957).
3. Phillips, C. Linear order and constituency. Linguist. Inq. 34, 3790 (2003).
4. Bemis, D.K. & Pylkknen, L. Basic linguistic composition recruits the left anterior
temporal lobe and left angular gyrus during both listening and reading. Cereb.
Cortex 23, 18591873 (2013).
5. Giraud, A.-L. & Poeppel, D. Cortical oscillations and speech processing: emerging
computational principles and operations. Nat. Neurosci. 15, 511517 (2012).
6. Sanders, L.D., Newport, E.L. & Neville, H.J. Segmenting nonsense: an event-related
potential index of perceived onsets in continuous speech. Nat. Neurosci. 5,
700703 (2002).
7. Bastiaansen, M., Magyari, L. & Hagoort, P. Syntactic unification operations are
reflected in oscillatory dynamics during on-line sentence comprehension. J. Cogn.
Neurosci. 22, 13331347 (2010).
8. Buiatti, M., Pea, M. & Dehaene-Lambertz, G. Investigating the neural correlates
of continuous speech computation with frequency-tagged neuroelectric responses.
Neuroimage 44, 509519 (2009).
 a r t ic l e s
9. Pallier, C., Devauchelle, A.-D. & Dehaene, S. Cortical representation of the constituent
structure of sentences. Proc. Natl. Acad. Sci. USA 108, 25222527 (2011).
10. Schroeder, C.E., Lakatos, P., Kajikawa, Y., Partan, S. & Puce, A. Neuronal oscillations
and visual amplification of speech. Trends Cogn. Sci. 12, 106113 (2008).
11. Buzski, G. Neural syntax: cell assemblies, synapsembles and readers. Neuron 68,
362385 (2010).
12. Bernacchia, A., Seo, H., Lee, D. & Wang, X.-J. A reservoir of time constants for
memory traces in cortical neurons. Nat. Neurosci. 14, 366372 (2011).
13. Lerner, Y., Honey, C.J., Silbert, L.J. & Hasson, U. Topographic mapping of a hierarchy
of temporal receptive windows using a narrated story. J. Neurosci. 31, 29062915
(2011).
14. Kiebel, S.J., Daunizeau, J. & Friston, K.J. A hierarchy of time-scales and the brain.
PLoS Comput. Biol. 4, e1000209 (2008).
15. Luo, H. & Poeppel, D. Phase patterns of neuronal responses reliably discriminate
speech in human auditory cortex. Neuron 54, 10011010 (2007).
16. Ding, N. & Simon, J.Z. Emergence of neural encoding of auditory objects while
listening to competing speakers. Proc. Natl. Acad. Sci. USA 109, 1185411859
(2012).
17. Zion Golumbic, E.M. et al. Mechanisms underlying selective neuronal tracking of
attended speech at a cocktail party. Neuron 77, 980991 (2013).
18. Peelle, J.E., Gross, J. & Davis, M.H. Phase-locked responses to speech in human
auditory cortex are enhanced during comprehension. Cereb. Cortex 23, 13781387
(2013).
19. Pasley, B.N. et al. Reconstructing speech from human auditory cortex. PLoS Biol.
10, e1001251 (2012).
20. Steinhauer, K., Alter, K. & Friederici, A.D. Brain potentials indicate immediate use
of prosodic cues in natural speech processing. Nat. Neurosci. 2, 191196
(1999).
2016 Nature America, Inc. All rights reserved.
21. Pea, M., Bonatti, L.L., Nespor, M. & Mehler, J. Signal-driven computations in
speech processing. Science 298, 604607 (2002).
22. Saffran, J.R., Aslin, R.N. & Newport, E.L. Statistical learning by 8-month-old
infants. Science 274, 19261928 (1996).
23. Ray, S. & Maunsell, J.H. Different origins of gamma rhythm and high-gamma activity
in macaque visual cortex. PLoS Biol. 9, e1000610 (2011).
24. Einevoll, G.T., Kayser, C., Logothetis, N.K. & Panzeri, S. Modeling and analysis of
local field potentials for studying the function of cortical circuits. Nat. Rev. Neurosci.
14, 770785 (2013).
25. Hagoort, P. & Indefrey, P. The neurobiology of language beyond single words.
Annu. Rev. Neurosci. 37, 347362 (2014).
26. Grodzinsky, Y. & Friederici, A.D. Neuroimaging of syntax and syntactic processing.
Curr. Opin. Neurobiol. 16, 240246 (2006).
27. Hickok, G. & Poeppel, D. The cortical organization of speech processing. Nat. Rev.
Neurosci. 8, 393402 (2007).
28. Friederici, A.D., Meyer, M. & von Cramon, D.Y. Auditory language comprehension:
an event-related fMRI study on the processing of syntactic and lexical information.
Brain Lang. 74, 289300 (2000).
29. Canolty, R.T. et al. High gamma power is phase-locked to theta oscillations in
human neocortex. Science 313, 16261628 (2006).
npg
164 VOLUME 19 | NUMBER 1 | JANUARY 2016 nature NEUROSCIENCE
30. Lakatos, P. et al. An oscillatory hierarchy controlling neuronal excitability and
stimulus processing in the auditory cortex. J. Neurophysiol. 94, 19041911
(2005).
31. Sirota, A., Csicsvari, J., Buhl, D. & Buzski, G. Communication between neocortex
and hippocampus during sleep in rodents. Proc. Natl. Acad. Sci. USA 100,
20652069 (2003).
32. Arnal, L.H. & Giraud, A.-L. Cortical oscillations and sensory predictions. Trends
Cogn. Sci. 16, 390398 (2012).
33. Poeppel, D., Idsardi, W.J. & van Wassenhove, V. Speech perception at the interface
of neurobiology and linguistics. Phil. Trans. R. Soc. Lond. B 363, 10711086
(2008).
34. Pea, M. & Melloni, L. Brain oscillations during spoken sentence processing.
J. Cogn. Neurosci. 24, 11491164 (2012).
35. Gross, J. et al. Speech rhythms and multiplexed oscillatory sensory coding in the
human brain. PLoS Biol. 11, e1001752 (2013).
36. Ding, N. & Simon, J.Z. Cortical entrainment to continuous speech: functional roles
and interpretations. Front. Hum. Neurosci. 8, 311 (2014).
37. Jackendoff, R. Foundations of Language: Brain, Meaning, Grammar, Evolution
(Oxford University Press, 2002).
38. Hagoort, P. On Broca, brain, and binding: a new framework. Trends Cogn. Sci. 9,
416423 (2005).
39. Cutler, A., Dahan, D. & van Donselaar, W. Prosody in the comprehension of spoken
language: a literature review. Lang. Speech 40, 141201 (1997).
40. Frazier, L., Carlson, K. & Clifton, C. Jr. Prosodic phrasing is central to language
comprehension. Trends Cogn. Sci. 10, 244249 (2006).
41. Singer, W. & Gray, C.M. Visual feature integration and the temporal correlation
hypothesis. Annu. Rev. Neurosci. 18, 555586 (1995).
42. Friederici, A.D. Towards a neural basis of auditory sentence processing. Trends
Cogn. Sci. 6, 7884 (2002).
43. Kutas, M. & Federmeier, K.D. Electrophysiology reveals semantic memory use in
language comprehension. Trends Cogn. Sci. 4, 463470 (2000).
44. Neville, H., Nicol, J.L., Barss, A., Forster, K.I. & Garrett, M.F. Syntactically based
sentence processing classes: evidence from event-related brain potentials. J. Cogn.
Neurosci. 3, 151165 (1991).
45. Lau, E.F., Phillips, C. & Poeppel, D. A cortical network for semantics: (de)constructing
the N400. Nat. Rev. Neurosci. 9, 920933 (2008).
46. Halgren, E. et al. N400-like magnetoencephalography responses modulated by
semantic context, word frequency and lexical class in sentences. Neuroimage 17,
11011116 (2002).
47. Van Petten, C. & Kutas, M. Interactions between sentence context and word
frequency in event-related brain potentials. Mem. Cognit. 18, 380393 (1990).
48. OConnell, R.G., Dockree, P.M. & Kelly, S.P. A supramodal accumulation-to-bound
signal that determines perceptual decisions in humans. Nat. Neurosci. 15,
17291735 (2012).
49. Koechlin, E., Ody, C. & Kouneiher, F. The architecture of cognitive control in the
human prefrontal cortex. Science 302, 11811185 (2003).
50. Nozaradan, S., Peretz, I., Missal, M. & Mouraux, A. Tagging the neuronal entrainment
to beat and meter. J. Neurosci. 31, 1023410240 (2011).
 ONLINE METHODS
Participants. 34 native listeners of Mandarin Chinese (1936 years old, mean
25 years old; 13 male) and 13 native listeners of American English (2246 years
old, mean 26 years old; 6 male) participated in the study. All Chinese listeners
received high school education in China and 26 of them also received college
education in China. None of the English listeners understood Chinese. All partici-
pants were right-handed51. Five experiments were run for Chinese listeners and
two experiments for English listeners. Each experiment included eight listeners
(except that the AMS experiment involved five listeners) and each listener par-
ticipated in at most two experiments. The number of listeners per experiment was
chosen based on previous MEG experiments on neural tracking of continuous
speech. The sample size for previous experiments was typically between three
and 12 (refs. 15,16), and the basic phenomenon reported here was replicated
in all the seven experiments of the study (N = 47 in total). The experimental
procedures were approved by the New York University Institutional Review
Board, and written informed consent was obtained from each participant before
the experiment.
Stimuli I: Chinese materials. All Chinese materials were constructed based on
an isochronous sequence of syllables. Even when the syllables were hierarchi-
cally grouped into linguistic constituents, no acoustic gaps were inserted between
constituents. All syllables were synthesized independently using the Neospeech
synthesizer (http://www.neospeech.com/, the male voice, Liang). The synthesized
2016 Nature America, Inc. All rights reserved.
syllables were 75354 ms in duration (mean duration 224 ms), and were adjusted
to 250 ms by truncation or padding silence at the end. The last 25 ms of each
syllable were smoothed by a cosine window.
Four-syllable sentences. 50 four-syllable sentences were constructed, in which
the first two syllables formed a noun phrase and the last two syllables formed a
verb phrase (Supplementary Table 1). The noun phrase could be composed of
either a single two-syllable noun or a one-syllable adjective followed by a one-
syllable noun. The verb phrase could be composed of either a two-syllable verb or
a one-syllable verb followed by a one-syllable noun object. In a normal trial, ten
sentences were sequentially played and no acoustic gaps were inserted between
sentences (Supplementary Fig. 1a). Due to the lack of phrasal and sentential
level prosodic cues, the sound intensity of the stimulus, characterized by the
sound envelope, only fluctuates at the syllabic rate but not at the phrasal or sen-
tential rate (Supplementary Fig. 2). An outlier trial was the same as a normal
trial except that the verb phrases in two sentences were exchanged, creating two
nonsense sentences with incompatible subjects and predicates (an example in
English would be new plans rub skin).
Four-syllable verb phrases. Two types of four-syllable verb phrases were
created. Type I verb phrase contained a one-syllable verb followed by a three-
npg
syllable noun phrase, which could be a compound noun or an adjective + noun
phrase (Supplementary Fig. 1b and Supplementary Table 1). Type II verb phrase
contained a two-syllable verb followed by a two-syllable noun (Supplementary
Fig. 1c, all phrases listed in Supplementary Table 1). 50 instances were created
for each type of verb phrases. In a normal trial, ten phrases of the same type were
sequentially presented. An outlier trial was the same as a normal trial except that
the verbs in two phrases were exchanged, creating two nonsense verb phrases
with incompatible verbs and objects (an example in English would be drink a
long walk).
Two-syllable phrases. The verb phrases (or the noun phrases) in the four-
syllable sentences were presented in a sequence (Supplementary Fig. 1d).
In a normal trial, 20 different phrases were played. In an outlier trial, one of
the 20 phrases was replaced by two random syllables that did not constitute a
sensible phrase.
Random syllabic sequences. The random syllabic sequences were generated
based on the four-syllable sentences. Each four-syllable sentence was trans-
formed into four random syllables using the following rule: the first syllable in
the sentence was replaced by the first syllable of a randomly chosen sentence.
The second syllable was replaced by the second syllable of another randomly
chosen sentence and the same for the third and the fourth syllables. This way, if
there were any consistent acoustic differences between the syllables at different
positions in a sentence, those acoustic differences were preserved in the random
syllabic sequences. Each normal trial contained 40 syllables. In outlier trials, four
consecutive syllables were replaced by a Chinese idiom.
doi:10.1038/nn.4186
Backward syllabic sequences. In normal trials, ten four-syllable sentences were
played but with all syllables being played backward in time. An outlier trial was
the same as a normal trial except that four consecutive syllables at a random
position were replaced by four random syllables that were not reversed in time.
Four-syllable idioms. 50 common four-syllable idioms were selected
(Supplementary Table 1), in which the first two syllables formed a noun phrase
and the last two syllables formed a verb phrase. In a normal trial ten sentences
were played. An outlier trial was the same as a normal trial except that the noun
phrases in two idioms were exchanged, creating two nonexistent and semanti-
cally nonsensical idioms.
Sentences with variable duration and syntactic structures. The sentence dura-
tion was varied between four and eight syllables. 40 sentences were constructed
for each duration, resulting in a total of 200 sentences (listed in Supplementary
Table 1). All 200 sentences were intermixed. In a normal trial, ten different
sentences were sequentially played without inserting any acoustic gap in
between sentences. In an outlier trial, one of the ten sentences was replaced by a
syntactically correct but semantically anomalous sentence. Examples of nonsense
sentences, translated into English, included ancient history is drinking tea and
take part in his portable hard drive.
Sentences with variable NP durations. All sentences consisted of a noun phrase
followed by a verb phrase (Supplementary Table 1). The noun phrase had three
syllables for half of the sentences (N = 45) and four syllables for the other half.
A three-syllable noun phrase was followed by either a four-syllable verb phrase
(N = 20) or a five-syllable verb phrase (N = 25). A four-syllable noun phrase was
followed by a three-syllable verb phrase (N = 20) or a four-syllable verb phrase
(N = 25). Sentences with different noun phrase durations and verb phrase dura-
tions were intermixed. In a normal trial 10 different sentences were played
sequentially, without inserting any acoustic gap between phrases or sentences.
In an outlier trial one sentence was replaced by a sentence with the same syntactic
structure but that was semantically anomalous.
AMS. Five sets of AMS were created. Each sentence consisted of three compo-
nents, C1, C2 and C3. Each component (C1, C2 or C3) was independently chosen
from three candidate syllables with equal probability. The grammar of the AMS
is illustrated in Supplementary Figure 4a. In the experiments, sentences were
played sequentially without any gap between sentences. Since all components
were chosen independently and each component was chosen from three syllables
with equal probability, all components were equally predictable regardless of its
position in a sequence. In other words, P(C1) = P(C2) = P(C3) = P(C2|C1) =
P(C3|C2) = P(C1|C3) = 1/3.
All Chinese syllables were synthesized independently and adjusted to 300 ms
by truncation or padding silence at the end. In each trial, 60 sentences were
played and no additional gap was inserted between sentences. Therefore, the
syllables were played at a constant rate of 3.33 Hz and the sentences were played
at a constant rate of 1.11 Hz. To make sure that neural encoding of the AMS was
not confounded by acoustic properties of a particular set of syllables, five sets of
AMS were created (Supplementary Table 1). No meaningful Chinese expressions
are embedded in the AMS sequences.
Stimuli II: English materials. All English materials were synthesized using the
MacinTalk Synthesizer (male voice Alex, in Mac OS X 10.7.5).
Four-syllable English sentences. 60 four-syllable English sentences were con-
structed (Supplementary Table 1), and each syllable was a monosyllabic word. All
sentences had the same syntactic structure: adjective/pronoun + noun + verb +
noun. Each syllable was synthesized independently, and all the synthesized
syllables (250347 ms in duration) were adjusted to 320 ms by padding silence
at the end or truncation. The offset of each syllable was smoothed by a 25-ms
cosine window. In each trial, 12 sentences were presented without any acous-
tic gap between them. In an outlier trial, 3 consecutive words from a random
position were replaced by three random words so that the corresponding
sentence(s) became ungrammatical.
Shuffled sequences. Shuffled sequences were constructed as an unintelligible
sound sequence that preserved the acoustic properties of the sentence sequences.
All syllables in the four-syllable English sentences were segmented into five over-
lapping slices. Each slice was 72 ms in duration and overlapped with neighboring
slices for 10 ms. The first 10 ms and the last 10 ms of each slice was smoothed by a
linear ramp, except for the onset of the first slice and the offset of the last slice.
nature NEUROSCIENCE
 A shuffled sentence was constructed by shuffling all slices at the same position
across the four-syllable sentences. For example, the first slice of the first syllable
in a given sentence was replaced by the first slice of the first syllable in a different
randomly chosen sentence. For another example, the third slice of the fourth
syllable in one sentence was replaced by the third slice of the fourth syllable in
another randomly chosen sentence. In a normal trial, 12 different shuffled sen-
tences were played sequentially, resulting in a trial that had the same duration
as a trial of four-syllable English sentences. In an outlier trial, four consecutive
shuffled syllables were replaced by four randomly chosen English words that did
not construct a sentence.
Markovian sentences. The pronunciation of an English syllable strongly
depends on its neighbors. To simulate more natural English, we also synthe-
sized English sentences based on an isochronous multi-syllabic acoustic chunk.
Every sentence was divided into three acoustic chunks that were roughly equal in
duration. Each acoustic chunk consisted of 12 monosyllabic or bisyllabic words
and was synthesized as a whole, independently of neighboring acoustic chunks.
All synthesized acoustic chunks (250364 ms in duration) were adjusted to 350 ms
by truncation or padding silence at the end. The offset of each chunk was
smoothed by a 25-ms cosine window.
Two types of Markov chain sentences were generated based on isochronous
sequences of acoustic chunks. In one type of Markovian sentences, called the
constant predictability sentences, each acoustic chunk was equally predictable
based on the preceding chunk, regardless of its position within a sentence.
2016 Nature America, Inc. All rights reserved.
The constant predictability sentences were generated based on the grammar
specified in Figure 3a and Supplementary Figure 1e. Listeners were familiar-
ized with the grammar and were able to write down the full grammar table
before participating in the experiment. In each trial, ten sentences were sepa-
rately generated based on the grammar and sequentially presented without any
acoustic gap between them.
The other type of Markovian sentences, called the predictable sentences, con-
sisted of a finite number of sentences (N = 25, Supplementary Table 1) that
were extensively repeated (1112 times) in a ~7-min block. In these sentences,
the second and the third acoustic chunks were uniquely determined by the first
chunk. In each trial, ten different sentences were played sequentially without any
acoustic gap between them.
Acoustic analysis. The intensity fluctuation of the sound stimulus is charac-
terized by its temporal envelope. To extract the temporal envelope, the sound
signal is first half-wave rectified and then downsampled to 200 Hz. The Discrete
Fourier Transform of the temporal envelope (without any windowing) is shown in
Figure 1 and Supplementary Figure 2.
npg
Experimental procedures. Seven experiments were run. Experiment 14
involved Chinese listeners listening to Chinese materials, experiment 5 involved
English listeners listening to Chinese materials, and experiment 6 involved
English listeners listening to English materials. Experiment 7 involved Chinese
listeners listening to AMS.
In all experiments except for experiment 5, listeners were instructed to detect
outlier trials. At the end of each trial, listeners had to report whether it was a
normal trial or an outlier trial via button press. Following the button press,
the next trial was played after a delay randomized between 800 and 1,400 ms.
In experiment 5, listeners performed a syllable counting task described below.
Behavioral results are reported in Supplementary Table 2.
Experiment 1. Four-syllable Chinese sentences, four-syllable idioms, random
syllabic sequences and backward syllabic sequences were presented in separate
blocks. The order of the blocks was counter balanced across listeners. Listeners
took breaks between blocks. In each block, 20 normal trials and ten outlier trials
were intermixed and presented in a random order.
Experiment 2. Four-syllable sentences, type I four-syllable verb phrases, type II
four-syllable verb phrases, two-syllable noun phrases, and two-syllable verb
phrases were presented in separate blocks. The order of the blocks was counter
balanced across listeners. Listeners took breaks between blocks. In each block,
20 normal trials and five outlier trials were intermixed and presented in a
random order.
Experiment 3. Sentences with variable durations and syntactic structures, as
described above, were played in an intermixed order. Listeners took a break every
25 trials. In total, 80 normal trials and 20 outlier trials were presented.
nature NEUROSCIENCE
Experiment 4. Sentences with variable NP durations, as described above,
were presented in a single block, counterbalanced with three other blocks that
presented language materials not analyzed here. In that block, 27 normal trials
and seven outlier trials were presented. The other three blocks presented other
language materials not analyzed here. The order of the blocks was counterbal-
anced across listeners.
Experiment 5. Trials consisting of four-syllable sentences, four-syllable idioms,
random syllabic sequences, and backward syllabic sequences were intermixed and
presented in a random order. 20 normal trials for each type of materials were pre-
sented. In each trial, the last 1 or 2 syllable was removed, each with 50% probability.
Listeners were instructed to count the number of syllables in each trial in a cyclic
way: 1, 2, 3, 4, 1, 2, 3, 4, 1, 2 The final count could only be 2 or 3 and the listeners
had to report whether it was 2 or 3 at the end of each trial via button press.
Experiment 6. Four-syllable English sentences, shuffled sequences, constant
predictability Markovian sentences, and predictable Markovian sentences were
presented in separate blocks. The order of the blocks was counterbalanced across
listeners. Listeners took breaks between blocks. In each block, 22 normal trials
and 8 outlier trials were intermixed and presented in a random order.
Experiment 7. The experiment involved the AMS and was divided into two
sessions. In the first session, ten trials were presented (two trials from each AMS
set; see the upper row in Supplementary Fig. 4b). In each trial, the last syllable was
removed with 50% probability. The listeners were told that the stimulus was only a
sequence of random syllables. They were asked to count the number of syllables in a
cyclic way: 1, 2, 1, 2, 1, 2 and report whether the final count was 1 or 2 at the end
of each trial via button press. Since each trial contained 179 or 180 rapidly presented
syllables, the listeners were not able to count accurately (mean performance 52
9.7%, not significantly above chance, P > 0.8, t test). However, the listeners were
asked to follow the rhythm and keep counting even when they lost count. After
the first session of the experiment was finished, the listeners were told about that
the general structure of the AMS and examples were given based on real Chinese
sentences. In the second session of the experiment, the listeners had to learn the
5 sets of AMS separately (lower row, Supplementary Fig. 4b). For each set of the
AMS, during training, the listeners listened to 20 sentences from the AMS set in a
sequence, with a 300-ms gap being inserted between sentences to facilitate learning.
Then, the listeners listened to two trials of sentences from the same AMS set, which
they also listened to in the first session (shown by symbol S in Supplementary
Fig. 4b). They had to do the same cyclical counting task. However, they were told
that the last count was 1 if the last sentence was incomplete and the last count was 2
if the last sentence was complete (mean performance 82 8.0%, significantly above
chance P < 0.2, t test). At the end of the two trials, the listeners had to report the
grammar of the AMS, i.e. which 3 syllables could be the first syllable of a sentence,
which three syllables could be the middle one, and which three syllables could
be the last one. The grammatical roles of 77 7.6% (mean s.e. across subjects)
syllables were reported correctly.
Neural recordings. Cortical neuromagnetic activity was recorded using a
157-channel whole-head MEG system (KIT) in a magnetically shielded room.
The MEG signals were sampled at 1 kHz, with a 200-Hz low-pass filter and a
60-Hz notch filter applied online and a 0.5-Hz high-pass filter applied offline
(time delay compensated). The environmental magnetic field was recorded using
three reference sensors and regressed out from the MEG signals using time-
shifted PCA52. Then, the MEG responses were further denoised using the blind
source separation technique, Denoising Source Separation (DSS)53. The MEG
responses were decomposed into DSS components using a set of linear spatial
filters, and the first 6 DSS components were retained for analysis and transformed
back to the sensor space. The DSS decomposes multi-channel MEG recordings
to extract neural response components that are consistent over trials and has
been demonstrated to be effective in denoising cortical responses to connected
speech18,54,55. The DSS was applied to more accurately estimate the strength of
neural activity phase-locked to the stimulus. Even when the DSS spatial filtering
process was omitted, for the r.m.s. response over all MEG sensors, the senten-
tial, phrasal, and syllabic responses in Figure 1 were still statistically significant
(P < 0.001, bootstrap).
Data analysis. Only the MEG responses to normal trials were analyzed.
Frequency domain analysis. In experiments 1, 2, 5 and 6, the linguistic
structures of different hierarchies were presented at unique and constant rates
doi:10.1038/nn.4186
 and neural tracking of those linguistic structures was analyzed in the frequency
domain. For each trial, to avoid the transient response to the acoustic onset of each
trial, the neural recordings were analyzed in a time window between the onset of
the second sentence (or the fifth syllable if the stimulus contained no sentential
structure) and the end of the trial. The single-trial responses were transformed into
the frequency domain using the discrete Fourier transform (DFT). For all Chinese
materials and the artificial Markovian language materials, nine sentences were ana-
lyzed in each trial, resulting in a frequency resolution of 1/9 of the sentential rate
(~0.11 Hz). For the English sentences and the shuffled sequences, the trials were
longer and the duration equivalent to 44 English syllables was analyzed, resulting
in a frequency resolution of 1/44 of the syllabic rate, that is, 0.071 Hz.
The response topography (Fig. 1c) showed the power of the DFT coefficients
at a given frequency and hemispheric lateralization was calculated by averaging
the response power over the sensors in each hemisphere (N = 54).
Given that the properties of the neural responses to linguistic structures and
background neural activity might vary in different frequency bands, to treat each
frequency band equally, a separate spatial filter was designed for every frequency
bin in the DFT output to optimally estimate the response strength. The linear
spatial filter was the DSS filter56. The output of the DSS filter was a weighted
summation over all MEG sensors, and the weights were optimized to extract
neural activity consistent over trials. In brief, if the DFT of the MEG response
averaged over trials is X(f) and the autocorrelation matrix of single-trial MEG
recordings is R(f), the spatial filter is w = R1(f)X(f) (see the appendix of ref. 56).
2016 Nature America, Inc. All rights reserved.
The spatial filter w is an 157 1 vector (for the 157 sensors), the same size as X(f),
and R(f) is a 157 157 matrix. The spatial filter could be viewed as a virtual sen-
sor that was optimized to record phase-locked neural activity at each frequency.
Power of the scalar output of the spatial filter, |XT(f)R1(f)X(f)|2, was the power
spectral density shown in the figures.
Time domain analysis. The response to each sentence was baseline corrected
based on the 100-ms period preceding the sentence onset, for each sensor.
To remove the neural response to the 4-Hz isochronous syllabic rhythm and focus
on the neural tracking of sentential/phrasal structures, we low-pass filtered the
neural response waveforms using a 0.5-s duration linear phase FIR filter (cut-off
3.5 Hz). The filter delay was compensated by filtering the neural signals twice,
once forward and once backward. When the response power at 4 Hz was extracted
separately by a Fourier analysis, it does not significantly change as a function of
sentence duration (P > 0.19, one-way ANOVA). The r.m.s. of the MEG responses
was calculated as the sum of response power (that is, square of the MEG response)
of all sensors, and the r.m.s. response was further low-pass filtered by a 0.5-s
duration linear phase FIR filter (cut-off 3.5 Hz, delay compensated).
A linear decoder was built to decode the duration of sentences. In the decoding
analysis, the multi-channel MEG responses were compressed to a single channel,
npg
i.e. the first DSS component, and the decoder solely relied on the time course of
the neural response. A 2.25-s response epoch was extracted for each sentence,
starting from the sentence onset. A leave-one-out cross-validation procedure
was employed to evaluate the decoders performance. Each time, the response
to one sentence was used as the testing response, and the responses to all other
sentences were treated as the training set. The training signals were averaged
for sentences of the same duration, creating a template for the response time
course for each sentence duration. The testing response was correlated with all
the templates and the category of the most correlated templates was the decoders
output. For example, if the testing response was most correlated with the template
for 5-syllable sentences, the decoders output would be that the testing response
was generated by a five-syllable sentence.
Statistical analysis and significance tests. For spectral peaks (Figs. 1 and 2), a
one-tailed paired t test was used to test if the neural response in a frequency bin
was significantly stronger than the average of the neighboring four frequency
bins (two bins on each side). Such a test was applied to all frequency bins below
5 Hz, and a FDR correction for multiple comparisons was applied. Except for the
analysis of the spectral peaks, two-tailed t tests were applied. For all the t tests
applied in this study, data from the two conditions had comparable variance
and showed no clear deviation from the normal distribution when checking the
histograms. If the t test was replaced by a bias-corrected and accelerated
bootstrap, all results remained significant.
In Figure 4, the s.e.m. over subjects was calculated using bias-corrected and
accelerated bootstrap57. In the bootstrap procedure, all the subjects were resampled
doi:10.1038/nn.4186
with replacement 2,000 times. The s.d. of the resampled results was taken as the
s.e.m. In Figure 4d, the statistical difference between the two curves, that is, the
three-syllable NP condition and the four-syllable NP condition, was also tested
using bootstrap. For each subject, the difference between the responses in these
two conditions was taken. At each time point, the response difference was resam-
pled with replacement 2,000 times across the eight subjects, and percentage of the
resampled differences being larger or smaller than 0 (the smaller value) was taken
as the significance level. A FDR correction was applied to the bootstrap results.
Code availability. The computer code used for the MEG analyses is available
upon request.
Neural Source Localization Using ECoG. ECoG participants. ECoG recordings
were obtained from five patients (three female; average 33.6 years old, range 1942
years old) diagnosed with pharmaco-resistant epilepsy and undergoing clinically
motivated subdural electrode recordings at the New York University Langone
Medical Center. Patients provided informed consent before participating in the
study in accordance with the Institutional Review Board at the New York University
Langone Medical Center. Three patients were right-handed, two were left-handed.
All patients were native English speakers (one of them was a bilingual Spanish/
English speaker), and all patients were left-hemisphere dominant for language.
ECoG recordings. Participants were implanted with 96179 platinum-iridium
clinical subdural grid or strip electrodes (three patients with a left-hemisphere
implant and two patients with a right hemisphere implant, additional depth elec-
trodes implanted for some patients but not analyzed). The electrode locations per
patient are shown in Supplementary Figure 5. Electrode localization followed previ-
ously described procedures58. In brief, for each patient we obtained pre-op and post-
op T1-weighted MRIs which were co-registered with each other and normalized to a
MNI-152 template, allowing the extraction of the electrode location in MNI space.
The ECoG signals were recorded with a Nicolet clinical amplifier at a sampling
rate of 512 Hz. The ECoG recordings were re-referenced to the grand average over
all electrodes (after removing artifact-laden or noisy channels). Electrodes from
different subjects were pooled per hemisphere, resulting in 385/261 electrodes
in the left/right hemispheres. High gamma activity was extracted by high-pass
filtering the ECoG signal above 70 Hz (with additional notch filters at 120 and
180 Hz). The energy envelope of high gamma activity was extracted by taking
the square of high-gamma response waveform.
ECoG procedures. Participants performed the same task as healthy subjects in
the MEG (Fig. 2e,f). In brief, they listened to a set of English sentences and control
stimuli in the first and second block. The control stimulus, that is, the shuffled
sequences, preserves the syllabic-level acoustic rhythm of English sentences but
contain no hierarchical linguistic structure. The procedure was the same as the
MEG experiment, except for a familiarization session in which the subjects listened
to individual sentences with visual feedback. 60 trials of sentences and control stim-
uli were played. The ECoG data from each electrode was analyzed separately and
converted to the frequency domain via DFT (frequency resolution 0.071 Hz).
A significant response at the syllabic, phrasal or sentential rate was reported if
the response power at the target frequency was stronger than the response power
averaged over neighboring frequency bins (0.5-Hz range above and below the
target frequency). The significance level for each electrode was first determined
based on a bootstrap procedure that randomly sampled the 60 trials 1,000 times
and then underwent FDR correction for multiple comparisons across all elec-
trodes in the same hemisphere.
A Supplementary Methods Checklist is available.
51. Oldfield, R.C. The assessment and analysis of handedness: the Edinburgh inventory.
Neuropsychologia 9, 97113 (1971).
52. de Cheveign, A. & Simon, J.Z. Denoising based on time-shift PCA. J. Neurosci.
Methods 165, 297305 (2007).
53. de Cheveign, A. & Simon, J.Z. Denoising based on spatial filtering. J. Neurosci.
Methods 171, 331339 (2008).
54. Ding, N. & Simon, J.Z. Adaptive temporal encoding leads to a background-insensitive
cortical representation of speech. J. Neurosci. 33, 57285735 (2013).
55. Ding, N. & Simon, J.Z. Neural coding of continuous speech in auditory cortex during
monaural and dichotic listening. J. Neurophysiol. 107, 7889 (2012).
56. Wang, Y. et al. Sensitivity to temporal modulation rate and spectral bandwidth in the
human auditory system: MEG evidence. J. Neurophysiol. 107, 20332041 (2012).
57. Efron, B. & Tibshirani, R. An Introduction to the Bootstrap (CRC press, 1993).
58. Yang, A.I. et al. Localization of dense intracranial electrode arrays using magnetic
resonance imaging. Neuroimage 63, 157165 (2012).
nature NEUROSCIENCE